השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז.  בצ'אט

Join Zoom Meeting
https://zoom.us/j/282415583

Speaker: Tomer Cohen

M.Sc. student under the supervision of Prof. Lior Wolf

Sunday, April 26^th, 2020 at 14:00

     Zoom

Bidirectional One-Shot Unsupervised Domain Mapping

Abstract

We study the problem of mapping between a domain A, in which there is a single training sample and a domain B, for which we have a richer training set. The method we present is able to perform this mapping in both directions. For example, we can transfer all MNIST images to the visual domain captured by a single SVHN image and transform the SVHN image to the domain of the MNIST images. Our method is based on employing one encoder and one decoder for each domain, without utilizing weight sharing. The autoencoder of the single sample domain is trained to match both this sample and the latent space of domain B. Our results demonstrate convincing mapping between domains, where either the source or the target domain are defined by a single sample, far surpassing existing solutions.

Join Zoom Meeting
https://proofpoint.zoom.us/j/607284105

Speaker: Hadar Cohen-Duwek

Ph.D. student under the supervision of Hedva Sptizer and Alex Bronstein

ZOOM Seminar

Wednesday, 22^nd April at 12:00

Visual system's mechanisms and computational models for Spatio-temporal visual phenomena

Abstract

The research goal was to offer mechanisms and computational models that predict several visual spatio-temporal phenomena, which are a subgroup of the "visual illusions” phenomena.  (Visual illusions enable us to identify underling neuronal mechanisms of the visual system).

We focused on the “watercolor illusion”, the positive and the negative "aftereffect illusions" and the “reversed phi illusion”. The first group of illusions that we focused on involve the "creation" of surfaces which are triggered by edges, in both spatial and spatiotemporal domains. This type of effects are termed as filling-in processes. The second type of effects is directional motion effects and retinal processes. We suggest that the connection between these two types of effects derives from their common mechanism’s components.

For the filling in effects, we suggest a computational model that is based on a Poisson equation as heat source. Our proposed computational model is able to predict most of the filling-in effects that derive from edges, such as the assimilative and non-assimilative watercolor illusions, and the positive and negative aftereffects.

For the motion effects, we developed a computational model supported by simulations, which for the first time leads to correct predictions of both the behavioral motion effects (phi and reverse-phi), and the relevant electrophysiological and anatomical findings. This has been achieved through the well-known neuronal response: the rebound response (or "Off response"). For testing this hypothesis, we proposed a psychophysical experiment on human subjects, which demonstrated the abolishing of the aftereffect phenomenon through a noise mask. The experimental results confirmed the theoretical predictions regarding the role of the rebound mechanism.  We furthermore suggest that the reverse-phi effect is only an epiphenomenon of the rebound response of the visual system. Our findings, thus, shed new light on the comprehensive role of the rebound response as a parsimonious spatiotemporal detector for motion and additional memory tasks, such as for stabilization and navigation.

השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם + מספר ת.ז.  בצ'אט

Join Zoom Meeting

https://zoom.us/j/397031628

Meeting ID: 397 031 628

Password: 046415

Speaker: Yuval Barkan

Ph.D. student under the supervision of Hedva Spitzer and Shmuel Einav

MONDAY, April 20^th, 2020 at 14:00

      ZOOM

How low level visual mechanism research contribute both to understanding enigmatic visual illusions and to computer vision mechanisms

Abstract

Computer vision, similar to the biological visual system, strive to interpret and interact with the visual physical world based on noisy, occluded, illuminated, partial and often inconsistent flow of information.

Since visual illusions are results of visual mechanism they can shed light on critical mechanisms of the visual system.

In this Ph.D. thesis, we have studied different aspects of the low-level visual system mechanisms that aid the visual system to perform complicated tasks such as color and intensity constancy, enhancements of intensity and contrast that enable segmentation and classification. Furthermore, we investigated the mechanisms that perform compensations to optical imperfections of the visual system such as longitudinal chromatic aberration. We have also investigated different visual illusions in order to understand and model their underlying mechanisms. As part of this research we have also discovered new visual illusions such as the color-dove illusions.

Based on the above computational models we have developed unique algorithms for various computer vision applications, for both natural and medical images such as companding HDR (High Dynamic Range) images. The building blocks of these computational models enabled us to develop a novel algorithm to present a CT HDR image on a standard dynamic range display without losing image diagnostic features.

Understanding these low-level mechanisms can contribute to modern deep learning and neural network algorithms. One of the challenges of modern neural networks is how images and how to perform effectively non-linear feedbacks (such as PredNet).  We believe the suggested mechanisms and models in this work can contribute not only to understand the actual visual mechanisms, but also to neural networks algorithms as part of their architecture and/or preprocessing layers.